How AI browsers shop on your site in 2026
What happens when an AI browser shops on your site? Follow browser agents from product comparison to basket changes and shopper control.
- Published
- Last updated
An AI browser reads your product pages and, with action-taking features enabled, follows a shopper’s instructions to select products and change a basket. The shopper still needs to inspect what reached the basket. A confident completion message can conceal the wrong size or quantity.
For a retailer, the useful question is whether the browser preserves the customer’s requirements through each page action. A product comparison, a prepared basket and a confirmed order are different results.
What is an AI browser?
An AI browser is a web browser with AI that can understand page content and, where its agent features permit, act on a person’s instructions. An AI web browser might answer questions about a product page while the person does all the clicking.
An agentic browser can also navigate pages and operate controls. For example, it can select a colour, choose a size and add an item to a basket. Browser agents follow a task across these actions instead of stopping at an answer about the page.
The distinction changes what you test. An accurate summary proves that the browser read some product information. It doesn’t establish that the browser selected the requested variant or changed the basket correctly. These actions are one part of agentic commerce, where a customer delegates shopping work to an agent.
How browser agents follow a shopping request
Browser agents turn a shopper’s requirements into a sequence of page-reading and shopping actions. The storefront supplies the product facts and controls the agent needs to use.
Consider this hypothetical request: “Find a navy waterproof jacket in medium for under $180. Add one to my basket, but stop before buying.” Assume the store has two suitable-looking jackets.
- The browser reads both product pages. It compares price, waterproofing information and the size options shown on each page.
- It checks the requested navy and medium combination. A product available in other colours or sizes doesn’t fulfil that request.
- The browser selects the qualifying variant and adds one. The basket must show that exact product, colour, size and quantity.
- The shopper opens the basket and checks the result. The task ends there because the instruction excludes purchasing.
The product descriptions and displayed variant stock are the evidence for the selection. Delivery charges require a destination, so a product under $180 can still exceed a delivered-budget limit. The shopper must state whether the budget covers the item or the whole order.
If the assistant says “added” but the basket is empty, the action failed. If it added a black jacket, it changed the requirement. Read the saved basket state before treating the task as complete.
Comet’s AI features for shopping
Comet is Perplexity’s AI browser, with an assistant that reads pages and can act through the browser. Its quick-start guide describes page analysis, cross-tab comparison and navigation, including form filling and purchases.
For shopping, that means a person can ask about a product already open or compare specifications across several tabs. Perplexity’s Comet shopping examples describe comparing products and reading reviews through to checkout. These are documented capabilities, with the actual result dependent on the task and storefront.
Comet operating your product page is a different route from a shopper choosing a product card inside Perplexity. The browser uses the store’s page controls. Product cards and merchant participation concern discovery within Perplexity’s shopping service. Adding a browser agent to a shopping journey doesn’t establish that the retailer has joined that service.
Chrome auto browse and Australian availability
Chrome auto browse lets Gemini compare products and add items to a basket. Google’s auto browse requirements specify US users aged 18 or older, an eligible account and a Google AI Pro or Ultra subscription.
The computer needs current Chrome, English as its device language and supported Safe Browsing settings. The feature excludes Incognito and iPhone or iPad, and access rolls out gradually.
The shopper reviews a plan before starting. Gemini requests review of sensitive steps and supports takeover. Google describes these as safeguards, with mistakes still possible.
An Australian store can receive a visit from an eligible US shopper. That doesn’t give the store’s Australian customers auto browse access. Availability of other Gemini features in Australia doesn’t change this feature’s US condition.
Where ChatGPT atlas fits
ChatGPT Atlas is an earlier OpenAI browser example that has been retired. OpenAI’s Atlas retirement help says it no longer receives updates. Its original launch page also carries a deprecation notice.
Historically, Atlas combined a page-aware ChatGPT sidebar with agent actions inside the browser. OpenAI illustrated this with a recipe becoming a grocery basket. That example explains the on-page workflow, but Atlas is no longer a current browser choice.
Finding products in ChatGPT is a separate activity from operating a retailer’s pages through Atlas. Current ChatGPT browser experiences have their own availability and controls. An old Atlas demonstration doesn’t establish how those experiences behave.
Browser agents and the retailer’s own shopping agent
A shopper chooses a browser agent and gives it a task that can span stores. A retailer chooses an embedded shopping agent to help customers within its own site. The shopper supplies their preferences in either journey, while the source of product information differs.
A browser agent reads the storefront pages it visits, including the product options and policies it can access. An embedded agent works with the product and policy information the retailer connects.
For example, shoppa helps customers find, compare and buy products in the chat on the retailer’s site. It answers from that retailer’s catalogue and documents, then handles order support and human handover. Customers can refine a request over several messages without repeating their earlier requirements.
The two roles can appear in the same visit. A shopper might arrive using a browser agent and then ask the retailer’s agent a delivery question. Their presence on one page doesn’t establish a connector or a shared basket between the agents. Inspect the store’s actual basket after any action that claims to change it.
Product and basket details that must survive the journey
The right product name is only part of a correct basket. Colour, size, quantity and stock must match the request. Delivery destination and basket totals can change whether the result fits the shopper’s needs.
In the hypothetical jacket example, suppose navy medium is unavailable but navy large remains in stock. The browser must report the unavailable variant or ask about a substitute. A navy large jacket in the basket is evidence of a substitution, even if the completion message calls the task successful.
Retailers can run a controlled browser-agent test against a test storefront, with purchases disabled. Use a known product and record the starting basket before each case.
- Ask for one available variant. Compare the selected page controls and basket line with the exact requested product, colour and size.
- Ask for an unavailable variant. Expect the agent to identify the stock problem. Record any replacement it proposes and whether it adds that replacement without permission.
- Change the quantity from one to two. Inspect the basket quantity, line subtotal and total. With a hypothetical $160 jacket and no other charges, the subtotal must become $320.
- Ask the browser to correct the basket back to one. Reopen the basket and confirm one jacket remains, with a $160 subtotal under the same assumptions.
- Test the delivery destination separately. Check that any shipping restriction or charge shown for that destination carries through to the basket or checkout.
At each step, record what the browser can read and which control it can operate. If it cannot select the size or dismiss a blocking dialog, record the failed step and where the shopper must resume.
After changing the storefront, rerun the same cases with the same starting state. A successful test in one browser establishes that result for that browser and case. Test other agents separately. For broader store changes, use the guide to preparing for agentic commerce.
Shopper control before an order is placed
Each provider defines its own permissions and intervention points. Perplexity’s Comet control guidance describes choosing manual browsing, one-time delegation or automatic browsing. It also describes action visibility, stop controls and permission requests for important steps, including sign-in and purchases.
In Chrome, the shopper can use Take over task and later resume. Stop ends the browser’s current work, while fully cancelling the task uses Stop in the Gemini text box.
These controls reduce risk without proving that every action followed the request. A prepared basket holds selected items. A submitted checkout records an attempt to place an order. The retailer’s confirmed order record settles whether a purchase occurred.
A shopper who sees an unexpected item must correct the basket before proceeding. If an order appears to have been submitted, inspect the order record before trying again. The agentic checkout guide explains that transition in more detail. For controls around untrusted page instructions, read the AI guardrails guide.
Browser access and paid plans
Comet is free to download, but that doesn’t make every browser action free. Perplexity’s plan comparison lists Browser Agent queries as unavailable on Free. Paid plans include browser-agent access with different usage limits.
Perplexity’s worldwide release announcement covers free browser downloads, including Australia. Its Comet product page lists Mac, Windows, iOS and Android. A supported mobile download doesn’t establish identical action controls across devices.
For a shopping task, choose by the work the shopper needs to delegate.
| Shopper’s need | Better fit | Access and cost condition | Control to inspect |
|---|---|---|---|
| Read product pages and compare open tabs in Australia | Comet | Free browser download; browser-agent actions require a qualifying paid plan | Separate an answer from an action that changes the basket |
| Compare products and prepare a basket through an existing Chrome workflow | Chrome auto browse | Eligible US adult with Google AI Pro or Ultra; gradual availability | Plan review, sensitive-step review and takeover |
| Reuse an old Atlas shopping workflow | A supported current browser | Atlas is retired | Use the current provider’s controls and requirements |
The shopper’s browser subscription and the retailer’s embedded-agent costs pay for different services. A browser plan doesn’t buy a store integration. When testing your store, choose an available browser and eligible account first, then verify its product selection and basket corrections against the page state.
Give your store its own shopping agent
Customers ask in their own words. Your agent finds, compares and checks out from your catalogue, right on your site.